iiRecord
Agentic AI Atlas · GAP-PERF-002: Session Compaction
page:docs-harness-features-backlog-gaps-performance-gap-perf-002a5c.ai
II.
Page overview

page:docs-harness-features-backlog-gaps-performance-gap-perf-002

Reference · live

GAP-PERF-002: Session Compaction overview

Inspect the raw attributes, linked wiki pages, and inbound or outbound graph edges for page:docs-harness-features-backlog-gaps-performance-gap-perf-002.

PageOutgoing · 0Incoming · 1

Attributes

nodeKind
Page
sourcePath
docs/harness-features-backlog/gaps/performance/GAP-PERF-002.md
sourceKind
repo-docs
title
GAP-PERF-002: Session Compaction
displayName
GAP-PERF-002: Session Compaction
slug
docs/harness-features-backlog/gaps/performance/gap-perf-002
articlePath
wiki/docs/harness-features-backlog/gaps/performance/GAP-PERF-002.md
article
# GAP-PERF-002: Session Compaction | Field | Value | |-------|-------| | Category | performance | | Priority | Critical | | Effort | XL | | Status | Partial | ## Description Implement session-level compaction that summarizes and condenses orchestration history, preventing context window overflow during long runs. ## Current State The compression module provides FNV-1a hash-based dedup and library file caching. This is token-level compression, not session-level compaction. The harness never summarizes or condenses conversation history. ## Target State Multiple compaction strategies: tool-output summarization, resolved effect pruning, iteration digest. Auto-compact trigger based on token budget thresholds. PreCompact/PostCompact hooks for custom behavior. ## Dependencies - [GAP-PROMPT-001](../prompt-engineering/GAP-PROMPT-001.md) -- prompt strata model for identifying compactable sections ## Key Files | Component | Path | |-----------|------| | Compression module | `packages/babysitter-sdk/src/compression/` | | Hook dispatcher | `packages/babysitter-sdk/src/hooks/dispatcher.ts` | | State cache | `packages/babysitter-sdk/src/runtime/replay/stateCache.ts` | ## Recommendation Phase 3 implementation. Auto-compact trigger at token budget threshold. At least 2 strategies: tool-output summarization and iteration digest. Integrate via on-iteration-end hook.
documents
[]

Outgoing edges

None.

Incoming edges

contains_page1